Uses AI safely with confidential and personal data

Safety · competency safety/handles-data-safely

Taught in: the Safety course

Draws on: Responsible use

Learning objectives

Claim	Why	Example
Before pasting, the learner asks who can see the prompt, how long it is kept, and whether it may train a model.	The answers differ between a consumer chat product, an enterprise plan and a local model, and the same paste is fine in one and a breach in another.	The learner checks that the company's approved tool has "no training on inputs" in its terms before pasting a customer list, and uses that tool instead of a personal account.
The learner removes or replaces personal data and secrets the task does not need before sharing the rest.	Most tasks work as well on redacted data, and what is not sent cannot leak.	To ask for help with a spreadsheet formula, the learner replaces the names column with "Person 1, Person 2" and keeps the numbers.
When in doubt, the learner treats the prompt like an email to an outside company.	The rule is easy to apply and matches how most organizations classify data.	The learner would not email the unreleased quarterly numbers to a vendor, so they do not paste them either, and ask the data owner instead.

Served by: Redacting a document before you paste it, What may go into an AI tool

Discloses AI use where the audience expects it (base)

Claim	Why	Example
The learner says that AI helped when the audience would judge the work differently knowing it.	Disclosure is about not misleading, and the test is what the reader would want to know.	A translated contract carries "machine-translated, reviewed by X", and a spell-checked email does not.
The learner follows the disclosure rule of the place the work goes to, and asks when there is none.	Journals, employers, schools and open-source projects have different rules, and breaking one costs more than the disclosure.	Before submitting to a conference, the learner reads the AI policy and adds the required statement.
Disclosure names what the AI did and what the person checked.	"Written with AI" tells the reader nothing about what to trust, while "drafted by AI, facts checked by me" does.	A commit message says which agent wrote the code and that the author ran and read the tests.

Served by: Introduction to the EU AI Act, What may go into an AI tool, Saying that AI helped, and crediting what it copied

Respects licenses and attribution in AI-assisted output (base)

Claim	Why	Example
The learner treats AI output that reproduces a recognizable source the same as copying that source.	A model can repeat text or code it saw in training, and the license of the original still applies to the copy.	A generated function that matches a well-known library almost word for word gets the library's attribution, or gets rewritten.
When adapting licensed material with AI help, the learner keeps the original's license terms and attribution.	The transformation does not remove the obligation, and the derived work carries it.	Rewriting a CC BY-SA tutorial with an assistant, the learner keeps the attribution and releases the result under the same license.
The learner knows the terms of the tool they use for who owns the output and what the tool may do with the input.	The terms differ between tools and plans, and the answer decides whether the output may go into a product.	Before shipping generated images in a brochure, the learner checks the image tool's commercial-use terms.

Served by: What may go into an AI tool, Saying that AI helped, and crediting what it copied

Alignment

Framework	Code	Asks	Objectives here
AI Fluency 4D (Dakan and Feller)	Diligence	Use AI responsibly, transparently and with accountability for the result	decides-what-to-share, discloses-ai-use, respects-licenses